Real Time Vehicle Country of Origin Classification Based on Computer Vision

Transcription

1 University of Zagreb Faculty of Transport and Traffic Sciences Real Time Vehicle Country of Origin Classification Based on Computer Vision Kristian Kovačić, Edouard Ivanjko, Sergio Varela* * LUNDS UNIVERSITET, SWEDEN edouard.ivanjko@fpz.hr 1

2 University of Zagreb Faculty of Transport and Traffic Sciences University of Zagreb, Croatia Established in faculties and 3 academies research staff members and students Faculty of Transport and Traffic Sciences Established in departments Cover all transport modes, logistics, ITS, aeronautics 100 research staff members / students Publisher of the journal PROMET Traffic & Transportation Cited in SCIE, TRIS, Geobase, FLUIDEX, and Scopus 2

3 University of Zagreb Faculty of Transport and Traffic Sciences Outline Introduction Problems and approaches Vehicle classification Vehicle detection and license plate recognition Vehicle detection speed up Experimental results Conclusion and future work 3

4 Introduction Faculty of Transport and Traffic Sciences - Computer Vision Group Developing algorithms for road traffic analysis based on computer vision Applications Traffic management Dynamic behaviour of a road traffic system derived from known parameters Traffic flow between nodes in a traffic network Driver information system Origin-Destination analysis of traffic on highways Computation of current and estimated OD matrices of a road traffic network Possibility to estimate the route of a traced vehicle 4

5 Problems and approaches Road traffic analysis Problems of manual measurement of traffic parameters Inaccurate data due to human error Impracticable to measure data 24/7 Measuring number of passed vehicles on complex intersections requires a large number of people for counting Increase need in human resources Impracticable to measure complex traffic parameters (vehicles queue, vehicle velocity, distance between vehicles) Sensors for measuring traffic parameters Pneumatic road tube sensors and piezoelectric sensors Inductive loops and magnetic sensors Radars, LIDARs Video cameras (color, IR, multi-spectral) 5

6 Problems and approaches Computer vision in traffic analysis Current commercial systems use one camera per lane Detection and tracking of vehicles Based on performing segmentation between objects of interest and noninterest objects using various image processing ARH Tattile methods (Fg/Bg image segmentation, optical flow, Haar method, Hough method) Estimation of vehicle trajectory Based on knowing vehicle location at certain time Describing vehicle movement by mathematical models which take into account vehicle dynamics Estimating next possible location (trajectory) of the vehicle 6

7 Problems and approaches OD matrix analysis Trajectory of moving vehicle through road traffic network (from node A to node B) Providing unique identification to each vehicle that is passing through road traffic network using automatic number (license) plate recognition Reduction of false positive/negative vehicles using additional statistical information given from origindestination (OD) matrix 7

8 Vehicle classification System architecture OpenCV framework for vehicle detection and localization CARMEN Freeflow SDK for license plate recognition (LPR) 8

9 Vehicle detection License plate recognition (LPR) Application objectives Detection of vehicles in video Tracking static vehicles (if vehicle stops to move) Vehicle license plate recognition for further traffic analysis Vehicle detection Pre-processing image imported from video with Gaussian blur filter Passing image through foreground / background image segmentation algorithm Finding contours which localize regions of detected vehicles 9

10 Vehicle detection Speed up Disadvantages of currently developed application Vehicle detection depends on license plate recognition High requirements for system resources (slow execution of algorithm due to sub-optimal approach) Optimization approach Executing algorithms on GPU as much as possible Adding support for CPU SIMD instructions to algorithms which are incapable to run on GPU Performing computations using multiple threads Parallelization of image processing algorithms 10

11 Experimental results Accuracy and execution time Accuracy Approach Analysis with sharpener filter Analysis without sharpener filter Total evaluation time [min] Real vehicle Count Corrected Vehicles Wrong Vehicles Correct / Real [%] % % Execution time Approach Avg time [ms] Contours for loop Processing time of an image with vehicle Min time [ms] Avg time [ms] Min time [ms] Single-thread Multi-thread

12 Experimental results Vehicle classification Extracted classification of vehicles by its country of origin Test video length - 30 [min] COUNTRY NUMBER OF VEHICLE RATIO [%] Germany Poland Austria Czech Republic Croatia Slovenia Turkey Slovakia Others Total

13 Experimental results Arised problems Overlapping vehicles cause false positive and false negative detections Environment conditions (sun reflection, rapid lighting changes), camera vibrations caused by strong wind or passing of large vehicles 13

14 Conclusion Future work Developed application has shown possibility of extracting a large number of information from video footage License plate number vehicle country of origin, vehicle trajectory, flow, number of vehicles, etc. One camera can be used for multiple lanes First results promising Further development of the application is currently in progress and it consists of following goals Estimation of vehicle trajectory on a road traffic network Detection and analysis of vehicle queue Determination of vehicle velocity Computation of origin-destination matrix of large road traffic network for purposes of traffic modelling 14

15 University of Zagreb Faculty of Transport and Traffic Sciences Acknowledgment This research has been partially supported by University of Zagreb grant 2013-ZUID-21, EU COST action TU1102 European Union from the European Regional Development Fund by the project IPA2007/HR/16IPO/ VISTA - Computer Vision Innovations for Safe Traffic Leading institution University of Zagreb, Faculty of electrical engineering and computing University of Zagreb, Faculty of Transport and Traffic Sciences 15

16 University of Zagreb Faculty of Transport and Traffic Sciences Real Time Vehicle Country of Origin Classification Based on Computer Vision Kristian Kovačić, Edouard Ivanjko, Sergio Varela* * LUNDS UNIVERSITET, SWEDEN edouard.ivanjko@fpz.hr 16